In　modern　high-performance　microprocessors,　branch　prediction　is　crucial　for　mitigating　performance　losses　due　to　pipeline　stalls.　Incorrect　predictions　result　in　CPU　instruction　pipeline　halts,　mainly　from　a　few　systematically　challenging　complex　branches　(Hard-To-Predict,　H2P)　where　current　techniques　lack　accuracy.　Traditional　TAGE-based　branch　prediction　struggles　with　noisy　complex　branches　and　scalability　issues.　As　application　scales　and　types　grow,　traditional　high-performance　processor　branch　prediction　faces　challenges.　This　paper　introduces　SS-CNN:　a　Convolutional　Neural　Networks　(CNN)-based　complex　branch　prediction　algorithm.　It　identifies　relevant　branches　in　branch　history　using　CNN　and　extracts　features　for　prediction.　Additionally,　attention　mechanisms　and　slice　structures　are　incorporated.　Optimization　enhances　the　model＇s　ability　to　extract　features　and　identify　relevant　branches,　improving　prediction　accuracy.　Experimental　results　show　SS-CNN　achieves　86.46%　accuracy　for　complex　branches,　a　1.01%　improvement　over　TAGE　SC　L.　SS-CNN　reduces　average　MPKI　by　4.18%　compared　to　TAGE　SC　L,　signifying　enhanced　accuracy,　fewer　interruptions,　increased　throughput,　and　lower　power　consumption,　improving　computer　architecture　performance.　©　2023　IEEE.